Download PDF.pdf

An intelligent agent has multiple goals and limited resources in a dynamic environment, requiring it to control its actions to achieve a high global utility. Two control models are the planning model, which plans actions ahead of time, and the reactive model, which reacts to environmental events as they unfold. The planning model has strengths such as achieving desirable global properties and conserving resources, but weaknesses like high computational cost and inflexibility. The reactive model is flexible but has weaknesses like infeasibility of anticipating all run-time contingencies and uncertain global effectiveness.

To address these issues, control modes and mechanisms are distinguished, with control modes being behaviors that emphasize sensitivity to run-time events and commitment to specific actions. A control mechanism should support different modes and smooth transitions among them. A theoretical space of control situations and modes is characterized, with dimensions of uncertainty and constraint imposed by goals, and sensitivity to run-time events and commitment to specific actions.

The spaces are superimposed to indicate that control modes in particular regions produce the best performance in corresponding regions of the situation space, assuming adequate monitoring and preparation resources. The domain of skiing is used to illustrate the concept, with different control modes suitable for different skiing situations, such as recreational skiing, practice turns, and downhill skiing.